YCY,
in the example you site what was the actual outcome? 1 or 2 mA
I usually spend a little amount of time proving to my self that the extraction is believable in the beginning of using a kit for the first time. For example the RF devices usually have somewhere in the model guide measured vs modeled. So if I take the exact same device and extract it I should get a similar performance. Usually this is the case but sometimes they differ. There are a few reasons why this might be, one is that there are settings in the capacitance/resistance look up table generation that when generated might not contain the right resolution to be able to resolve certain geometries correctly. Typically an over estimation is the result in this case, but it could go the other way. The goal is always to find a reasonable path back to a measured result, because modelling is mostly, but not always accurate.
Step 1: check the accuracy of your extractor against measured data.
Check that your extractor follows the correct behavior when for example your adjust the pcell so that your source/drain are pushed apart. The extractor should follow a logical decrease in the source to drain capacitance depending on distance. If not then step 1 might be a coincidence. Get a RFdevice or two to match and your confidence in the tool should increase.
If you believe that it all makes sense then you may want to construct a new model of prelayout devices using a wrapper of some sort to try to predict the post layout. Again believing in the extractor come up with a scalable model for this. Why the foundry can't do this For regular devices I don't know, but that is what it is.
Step 2: try to estimate your parasitics at pre layout
Finally be discriminating of results that don't make sense either good or bad. Most errors occur from believing results to good to be true. Try look at measured results as a a way to refine your simulation test bench.
These are just a few points obvious or not which may help you get more predictable results.
http://rfcooltools.com